Estimating dynamic properties of objects from appearance

To interact with objects effectively, a robot can use model-based or model-free control approaches. The superior performance typical of model-based control comes at the cost of developing or learning an accurate model of the system to be controlled. In this paper, we suggest an approach that generates models for novel objects based on visual features of those objects. These models can then be used for anticipatory control. We demonstrate this approach by replicating an infant experiment on a pneumatic humanoid robot. Infants seem to use visual information to estimate the mass of rods, and when they are presented a rod with an unexpected length-to-mass relationship, infants produce a large overcompensating arm movement when compared to an object with an expected mass. Our replication shows that the visual model-based control approach qualitatively replicates the behavior observed in the infant experiment, whereas a popular model-free approach, PID control, does not.